Dynamic Cue Combination in Distributional Population Code Networks

In this chapter, we investigate how hierarchical networks of neural populations might code stable representations of transient stimulus variables along with the uncertainty about them, while allowing consistent semantics of inputs and outputs at every level of the neural hierarchy. One effective paradigm to study this question is the integration of information from multiple sensory sources. Significant behavioral and neurophysiological evidence substantiates the notion that when separate sensorymodalities receive correlated inputs about an external stimulus, the information is often combined to produce a coherent, unitary percept of the world (Stein & Meredith, 1993). This ability of neural systems to combine information from multiple sources affords considerable, ecologically relevant advantages. For instance, it has been shown to attenuate the effects of input variability owing to noise, thereby reducing the overall uncertainty in the final estimate of the stimulus variable. Sensory cue integration and sensorimotor combinations involve many important and challenging statistical issues; for example, the different sources of information are not equally reliable, and their reliability can change with changing stimulus conditions and task demands. Bayesian probability theory has provided a normative framework for predicting how sensory systems might integrate information optimally, to make perceptual inferences about the environment from noisy sensory data. These predictions have been quantitatively tested using psychophysical experiments that examine what computations are performed when reliability of the cues changes over time. Results from widely ranging studies of human cue integration are consistent in that subjects behave in a manner that takes the uncertainty in sensory inputs into account, to form statistically optimal estimates of stimulus variables (see Knill & Pouget, 2004, for a review). Several different computational strategies have been proposed andvalidated; some examples are linearweighted averaging (Hillis, Watt, Landy, & Banks, 2004; Jacobs, 1999; Knill & Saunders, 2003; van Beers, Sittig, & van der Gon, 1999), extensions thereof (Landy, Maloney, Johnston, & Young, 1995), multiplicative interactions, and fully probabilistic linear and nonlinear Bayesian inference (Knill, 2003). Regardless of the exact inference strategy, the common theme in all these studies is that sensory uncertainty determines the optimal weighting scheme for combining sensory inputs for perceptual judgments. Similarly, studies also demonstrate that sensory and motor uncertainties determine how sensory signals should be transformed and used to plan actions and guide behavior. An example of a commonly used paradigm involves manipulating continuous visual feedback from the hand to control pointingmovements. Subjects are able to compensate for experimentally induced changes